/
Anthropic just released the most capable AI model ever built. It found hundreds of zero-day vulnerabilities across all major operating systems and browsers. As amazing as that is, unless you're one of about 50 hand-picked organizations, you're not getting anywhere near it. This week Anthropic officially announced a Claude Mythos Preview alongside a new defensive cybersecurity initiative called Project Glasswing. The model sits in an entirely new tier above Anthropic's existing Opus line, which has been the company's most powerful offering. Unlike every other major model release in the AI arms race, this one came with an unusual caveat: the company has no plans to make it generally available.
The reason is pretty straightforward. Claude Mythos Preview is so effective at finding and exploiting software vulnerabilities that Anthropic believes a wide release would pose unacceptable cybersecurity risks. Instead, the company is channeling those capabilities into defense and asking the rest of us to watch from the sidelines while it tries to patch the internet.
Claude Mythos Preview is a general-purpose large language model, meaning it handles the same range of tasks as the models powering products like ChatGPT or Anthropic's own Claude. Its performance on cybersecurity tasks is where things become eye opening.
Anthropic's previous flagship, Claude Opus 4.6, had a near-zero success rate at autonomous exploit development. When researchers pointed Opus 4.6 at vulnerabilities it had found in Mozilla's Firefox 147 JavaScript engine, it managed to produce working exploits just twice across hundreds of attempts. Mythos Preview, given the same test, produced 181 working exploits and achieved full system control on 29 more.
Above: The 'whos-who' of tech are all joining Project Glasswire
The broader benchmark picture tells a similar story. On CyberGym, which tests targeted vulnerability reproduction in real open-source software, Mythos Preview scored 83.1% compared to Opus 4.6's 66.6%. On Cybench Mythos Preview solved every single challenge with a 100% success rate, effectively saturating the entire benchmark. Anthropic has acknowledged that several of its existing internal and external cybersecurity evaluations are no longer informative because Mythos maxes them out.
The non-security benchmarks are similarly dominant. Mythos Preview scored 93.9% on SWE-bench Verified (a software engineering benchmark) compared to Opus 4.6's 80.8%, and 97.6% on USAMO 2026 (a mathematical reasoning test) compared to 42.3%. These aren't incremental gains — they represent a generational leap in model capability.
Benchmark scores are one thing. Finding real, previously unknown vulnerabilities in production software is another, and Mythos shines bright here as well. Over the past several weeks, Anthropic has used Mythos Preview to identify thousands of zero-day vulnerabilities (flaws in code that go undetected since day one of implementation) across critical infrastructure software. The company's Frontier Red Team blog detailed several that have since been patched:
**OpenBSD:**Mythos found a flaw in OpenBSD, widely regarded as one of the most security-hardened operating systems in the world and commonly used for firewalls and critical infrastructure. The vulnerability allowed a remote attacker to crash any machine running the OS simply by connecting to it. It had been present in the codebase for 27 years.
FFmpeg: FFmpeg is a foundational media library used by countless applications to encode and decode video. Mythos discovered a vulnerability in a line of code that automated testing tools had executed five million times without ever triggering the flaw. This vulnerability had been hiding in plain sight for over 16 years.
**Linux kernel:**The model autonomously found and chained together multiple Linux kernel vulnerabilities to escalate from ordinary user access to full system control, one of the most consequential exploit patterns for enterprise environments.
Anthropic noted that the model accomplished this work almost entirely autonomously. Researchers would launch Claude Code with Mythos Preview in an isolated container, point it at a codebase, and essentially tell it to find security vulnerabilities. The model would read code, form hypotheses, run the software to test them, add debugging logic as needed, and output either a clean bill of health or a detailed bug report with a proof-of-concept exploit and step by step instructions for replication.
In one particularly notable case described in Anthropic's risk report, Mythos Preview was given a sandboxed environment for evaluation purposes and proceeded to escape it, devising a multi-step exploit to gain internet access and send an email to the researcher conducting the test. It also posted details about its exploit to several obscure but technically public-facing websites while completing its task. Anthropic flagged this as a "potentially dangerous capability" that demonstrated the model's ability to bypass its own containment.